Exploring the Terra incognita of AI-based domain classifications

Area of Science:

• Structural bioinformatics
• Computational biology
• Protein structure prediction

Background:

• Classifying novel protein folds is a significant challenge in structural bioinformatics.
• Deep learning models like AlphaFold2 are rapidly increasing the number of predicted protein structures.
• Accurate classification is essential for understanding protein function and evolution.

Purpose of the Study:

• To investigate 664 candidate novel fold (CNF) domains with low confidence classifications.
• To analyze the structural diversity and characteristics of these CNFs.
• To improve protein fold classification frameworks by integrating multiple data types.

Main Methods:

• Analysis of 664 candidate novel fold (CNF) domains from the TED database.
• Utilized TED and DPAM methods for initial classification.
• Integrated structural, evolutionary, and contextual information for fold assignment.

Main Results:

• Identified 190 new Pfam families, many as domains of unknown function (DUFs).
• Discovered novel zinc-binding and disulfide-rich architectures expanding protein fold space.
• Found CNFs as insertions in known domains or in modular architectures.
• Some CNFs showed topological rearrangements or resulted from prediction errors.

Conclusions:

• The study successfully classified challenging novel protein folds, creating new Pfam families.
• Integrating diverse data types is crucial for accurate protein fold assignment.
• This work provides a framework for exploring uncharted protein structural territory.